Custom AI processors do seem to be the flavour of the month. We have this IBM AIU, which is a custom ASIC optimised for Watson, and we also have the Chiplet Cloud announced, in which you assemble your LLM-specific ASIC from standard chiplets, to cut development time and cost.
The general processing unit or GPU began its life as just such a dedicated ASIC, the graphics processing unit. How long before we see the GIU - general intelligence unit - hijacked as a marketing thing, even though it just runs arbitrary LLMs and does nothing like general intelligence?